Great challenges with few subjects: statistical strategies for neuroscientists

Neuroscience is one of the most heavily experimental fields of biological and medical research. As such, statistical approaches have traditionally focused on testing specific predictions based upon well-focused hypotheses. However, neuroscience data are often derived from repeated measurements and stimulus type presentations with a limited number of subjects, some of which may have incomplete data per subject. Here we provide an introduction to a group of diverse and powerful statistical approaches, which we term the '5 Ms', which have been successfully used in other fields of biological research facing similar constraints. Specifically, we detail how M1: meta-analysis can combine, reconcile, and analyse between- and within-study results, M2: mixed-effects modelling is beneficial through replacing statistical tests involving pseudoreplication, M3: multiple imputation may be used to account for the biases caused by missing data arising from incomplete experimental protocols, and M4: model averaging from information-theoretic approaches allows to discriminate among alternative functional hypotheses. We also provide a brief introduction to Bayesian statistics using M5: Markov chain Monte Carlo (MCMC). Taken together, these approaches provide neuroscientists with a robust statistical toolbox containing elements that alleviate some of the analytical constraints generated by limited sample sizes, repeated subject use, and incomplete replicates of experimental manipulation.